Have you analyzed the stability of TimeGPT, i.e. when one new observation is added to the dataset and the forecast is updated, does the forecast change a lot? I'm doing a 36 month forecast and I wanted to check the accuracy and stability so I did a cross validation. In one case when I moved forward one month the forecast changed dramatically even though the first forecast was extremely accurate (the first point forecast of the green line was 0.9% from the actual value). I'm using 100 steps in finetune and I'm using timegpt-1, I tried the long horizon model and I got similar behavior.

regarding this question, it seems that the target variable is very seasonal, with some periods of high values followed by some periods of relatively low values. However, can you privately share with us a reproducible example? I tried to replicate the 36 month forecast using a rolling origin of 1 month, but didn't found your issue